Computing Point-of-View: Modeling and Simulating Judgments of Taste
|
|  Yüklə 8.55 Mb.
|
2.4 Model generalizationOnce a person’s explicit textual traces are acquired from her everyday texts, those traces seed the inference of a generalized person model. The leap from textual traces to generalized model can be idealized as the leap from discrete data points to a comprehensive model—meaning that a reaction can be calculated for virtually any input conceivable in a realm. A goal for the generalized person model is to approach the sophistication and comprehensiveness of a person’s perspective. To achieve generalization, three heuristic approaches are tried—spreading activation (over cultural topology and metadata hierarchies), analogy, and imprimer model supplementation. This rest of this section 1) argues that the generalization process is convergent and exhibits truth maintenance properties; and presents scenarios for generalization via 2) spreading activation, 3) analogy, and 4) impriming. § Convergence. Generalization from fragmentary textual traces should see convergence because a person’s expression of those textual traces was a non-arbitrary emanation of that person’s perspective. We define convergence as the observation that each new textual trace add to the agglomerating person model offers less and less information not already known to the model. Convergence would be important to observe when trying to create a comprehensive model of a person’s taste because it signals that the model is nearing circumscription of the person’s taste. We opine on some reasons for believing in the coherency and consistency of perspective. First, human memory and cognitive function are shaped by the principle of economy. Given that physical memory is resource constrained, mental structures optimize for maximum utility—e.g. dream-work can be regarded as a garbage collection process consolidating recent relevant experiences and rejecting extraneous aspects; creativity means re-appropriating knowledge tied to one context, into other unexpected contexts, thus the knowledge which tends to persist are those with overloaded utility, versatile enough to solve problems under a variety of contexts. Second, competency in the social world motivates a need to communicate oneself, and consistency of perspective facilitates one’s social intelligibility. That people compose themselves around intelligible personae is consistent with Goffman’s 1959 thesis that we interact with the world as a performer acts through various masks. This line of reasoning also supports a belief that person’s coherencies should be captured in a mined cultural topology because personae are the product of social and cultural negotiation. Third, tastes are intrinsically organized to minimize dissonance. Individuals’ consumptive choices tend to cohere around their design for a life-style—clusters of goods having common implications to lifestyle are known as ‘Diderot Unities’ (McCracken 1988) and ‘consumption constellations’ (Solomon & Assael 1987). Cognitive dissonance (Festinger 1957), the unpleasant experience of mental self-conflict, leads persons to re-organize their possessions and their identity to eliminate glaring dissonance. Truth maintenance. Truth maintenance (Doyle 1980) is the detection and elimination of intra-system contradictions. We argue that in the context of model generalization, methods like spreading activation has a truth maintenance effect because nodes may be activated from multiple other nodes (evidence corroboration). Usually, spreading activation is thought of only as a mechanism for semantic expansion, but in a densely connected network, spreading activation also maintains truth—the strength of that property increasing proportionally with the density of connections in the activation network’s topology. The following example compares spreading activation in a sparse network versus in a dense network. Suppose there is a semantic network whose connections are sparse, and suppose activation is spread outward from a seed set of nodes. In this case, it is unlikely that activations seeded from different nodes will cross paths one-hop away or two-hops away. This has the effect of semantic expansion, but does not have a truth maintenance property. Suppose now that there is a semantic fabric with very dense connections. Suppose again that spreading activation proceeds from a seed set of nodes. In this case, what is different is that activations seeded from different nodes will cross paths almost immediately, providing checks and balances on the information embodied in each. Two-hops away, it is likely that all the seed nodes’ activations will have crossed. This crossing action, we argue, has the effect of truth maintenance—if two crossing activations are both strong, their activations are additive, and the node at which the crossing takes place is dually corroborated; if in the case of the food fabric, a positive activation crosses a negative activation, their effect is to contradict each other, resulting in a neutral (non) activation. Finally, we note that semantic fabrics present much more ample opportunity for corroboration and contradiction than did the sparse logical truth maintenance and contradiction detection features in previous user modeling systems such as GUMS (Finin & Drager 1986) and UMT (Brajnik & Tasso 1994). § Spreading activation. Collins & Loftus (1975) originally posed spreading activation as a psychological theory of memory. The rationale between that associative reasoning spreads across nodes of semantic memories with some energy, which decays proportionally with the number of steps traversed. Salton & Buckley (1988) affirmed the usefulness of spreading activation as a heuristic technique for semantic expansion in information retrieval tasks. They also articulated further techniques such as a system for assigning discounts. Each time activation traversed a link, the activation a was discounted by some coefficient, resulting in e.g., 0.5a activation. Furthermore, nodes with a large number of outgoing links incurred a fan-out discount (sometimes called a branching factor discount). Our use of spreading activation over topic hierarchies and over cultural topology follows this standard discount scheme—applying discounts to spreading and to fan-out. Next, we walk through two scenario of use for spreading activation. In the realms of attitudes and humor, activations seeded by textual traces are spread along the lines of implicit topic hierarchies. Consider that an individual’s attitude toward the topic of ‘water conservation’ is pleasurable-arousing-dominant (+1.0,+1.0,+1.0), and consider that the same individual’s attitudes toward the topics ‘environmental protection’, and ‘recycling’ are unknown. Generalizing from the known attitude for ‘water conservation’, activation is spread along the edge supertopicOf(‘environmental protection’, ‘water conservation’), which is gotten from folksonomies such as DMOZ. Because activation flows ‘upstream’ from topic to super-topic, an upstream discount of 0.75 is applied, resulting in ‘environmental protection’ (+0.75,+0.75,+0.75). For simplicity, separate tallies of PAD and uncertainty are not maintained and uncertainty is folded into PAD. To make this simplifying assumption, PAD values must range (-1.0 to +1.0), and definitely not (0.0 to +1.0). Next, assuming that ‘recycling’ is the only sub-topic of ‘water conversation’, with a downstream discount of 0.5, the relaxation result would be ‘recycling’ (+0.5,+0.5,+0.5). Topics with multiple sub-topics impart their affection downstream with an additional ‘fan-out’ discount, which is heuristically set to be inversely proportional to the log of its number of sub-topics. The mechanism of spreading activation continues propagating activation from topic up to supertopic to super-supertopic, and from topic to subtopic to sub-subtopic, each time applying the appropriate discount, until the post-discount PAD value is negligibly small (neutral). In the realm of cultural taste, a person’s textual traces are relaxed over the cultural topology, resulting in a taste ethos. Textual traces are presumed to be a fragmentary expression of a more profound but ineffable ethos (character). Spreading activation over the semantic fabric of taste should converge upon the ethos. Suppose that A, B, and C are three nodes with strong mutual affinities, forming a clique. Supposing then that only A and B are known in the textual trace, the result of spreading activation is to implicate C also into the ethos of the individual, as both activations from A and B will corroborate C. Activation is also mediated by identity hubs, and taste cliques. If item A has a strong affinity to identity X, and X has strong affinity to items A, B, C, and D, then when activation spreads into X, all strong members of X are together pulled into the gathering ethos. An individual’s ethos is more likely to stumble into a few focused identity hubs (thus, not threatened by fan-out) and pick up all their member spokes than to be constituted by nodes that are not interconnected themselves by semantic mediators. This observation implies that spreading activation in the presence of semantic mediators acquires a hierarchical quality. § Analogy. In attitude model generalization, analogies are automatically made, mapping topics in the model to new topics outside the model, and propagating PAD values to those new topics. While spreading activation along topic hierarchies is often a context-preserving act (e.g. ‘environmental protection’ and ‘water conservation’) analogy is more often a cross-context act (e.g. ‘war’ and ‘pollution’). As such, analogy is definitely a heuristic technique for generalization. Analogy for the purposes of this thesis is defined as a relationship between concepts with similar attributes, though not usually sharing a direct taxonomic parent. Structure mapping (Gentner 1983; Fauconnier & Turner 2002) is a way to find analogy, and this method is operationalized in the ConceptNet commonsense reasoning system. For example, suppose an environmentalist’s textual traces are asked to react to the concept of ‘war’. Feeding ‘war’ into ConceptNet’s get_analogous_concepts() function results in a rank-ordered list of analogous concepts, among which are some topics known in the environmentalist’s textual traces. Actual output from ConceptNet is shown below, edited for legibility. [war is like storm] the concepts share: ==PropertyOf==> bad ==PropertyOf==> violent ==PropertyOf==> dangerous ==CapableOf==> destroy property [war is like pollution] the concepts share: ==PropertyOf==> evil ==CapableOf==> kill ==CapableOfReceivingAction==> cause ==CapableOfReceivingAction==> stop Based on these shared attributes, the topic ‘war’ inherits the environmentalist’s attitudes about ‘storm’ and ‘pollution’, but each time with a 0.5 uncertainty discount, just as in spreading activation. ConceptNet is capable of resolving not only keyword topics, but also second-order topics such as ‘eat burger’, ‘open door’, etc. Its capabilities are surveyed in Chapter 3. § Impriming. Minsky (forthcoming) introduced the notion of ‘imprimer’ as a mentor or confidant that one forms an attachment to, and engages in mimesis of. Minsky’s notion is supported by similar notions in the psychology literature, such as Freud’s (1915) theory of ‘introjection’—children unconsciously emulating their parents’ values—and Ogden’s (1979) theory of ‘projective identification’. According to Minsky, parents imprime their children, an advisor imprimes a student, and even fictional characters and cults can imprime. Minsky’s litmus test for an imprimer is one who can stir self-conscious emotions such as pride and embarrassment within oneself. In attitude modeling, imprimer relations are thus detected by looking for co-occurrences of self-conscious affect with potential impriming entities such as persons and cultures. Once identified though, corpora must be assembled for imprimers in order to model their attitudes. This unfortunately is still a supervised process. In generalization, imprimer’s models are taken to supplement the person’s own model in parts of the topic space where the individual’s model offers no insight. It cannot be emphasized enough that generalization via imprimers is experimental, heuristic, and is motivated by a desire to ultimately create perspective-based applications that can react to any input. Such applications are ultimately fail-soft because they are vehicles for provocation and have other means of verifiability. When heuristics like imprimers produce wrong inferences, which they will, the results are not catastrophic. Interestingly, the effect of imprimers on attitude prediction was measured in an evaluation of attitude modeling, and will be presented in Chapter 4. This section characterized the generalization process as convergent and truth-maintaining, and provided details on three generalization mechanisms of spreading activation, analogy, and impriming. Next, the final section focuses on issues in model application. |